Nuqta-Arabic-OCR 0.8B GGUF: Arabic OCR on phones and laptops

This repo holds the GGUF build of Nuqta-Arabic-OCR-0.8B for llama.cpp. It reads printed Arabic pages, copies the tashkeel that is printed and almost never adds any. It runs fully offline in about 1.6 GB of RAM, on Android, iOS, Mac, Windows and Linux, on CPU or GPU.

Website and guides: https://salahalhaismawi.github.io/nuqta-ocr/ · run Arabic OCR offline

File	Size	Use
`nuqta-ocr-0.8b-Q4_K_M.gguf`	560 MB	phones (recommended)
`nuqta-ocr-0.8b-Q8_0.gguf`	873 MB	laptops and desktops, closest to the original
`nuqta-ocr-0.8b-mmproj-f16.gguf`	205 MB	vision encoder, always needed next to the model

Accuracy against the full model

We scored Q4_K_M (llama.cpp, CPU) against the bf16 model (vLLM) with the same prompts, image sizes and scoring: greedy decoding, KITAB-Bench normalisation, character error rate (lower is better), on the same pages.

KITAB-Bench set (first 40 images)	bf16	Q4_K_M GGUF
Hindawi (book pages with tashkeel, at the recommended input size)	0.031	0.028
PATS-OCR (printed lines)	0.009	0.014
SynthesizeAR (synthetic lines)	0.041	0.038
Adab (poetry word crops)	0.275	0.291

All differences are within page-to-page noise. On the full benchmarks the 0.8B model scores KITAB-Bench CER 0.117 and Misraj-DocOCR MARS 77.1: see the main model card.

Run it

llama-mtmd-cli -m nuqta-ocr-0.8b-Q4_K_M.gguf --mmproj nuqta-ocr-0.8b-mmproj-f16.gguf \
  --image page.png --temp 0 -c 8192 --image-max-tokens 4096 \
  -p "Transcribe the Arabic text in this image exactly as written. Include only the diacritics (tashkeel) that are visible in the image."

Or serve an OpenAI-compatible API (one image plus the prompt per request):

llama-server -m nuqta-ocr-0.8b-Q4_K_M.gguf --mmproj nuqta-ocr-0.8b-mmproj-f16.gguf \
  -c 8192 --image-max-tokens 4096 --temp 0 --port 8080

Any app built on a recent llama.cpp with multimodal support (mtmd, Qwen3.5 architecture) can load these two files. Tested with llama.cpp commit 5aeac37 (October 2026).

Three settings that matter

Use the recommended input size: enlarge full pages to about 2.4 megapixels. The model was trained at that size, and a small page image costs real accuracy (on a 612×792 book page the reading went from many wrong words to almost perfect). The rule we use: if the image is page-shaped (long side ≤ 2× short side, short side ≥ 450 px), scale it to 2.4 MP; leave line crops as they are; never go above 2.4 MP. On a slow phone, 1.2 MP is a reasonable compromise: about twice as fast, with a few more word errors.
Pass --image-max-tokens 4096, so llama.cpp doesn't shrink the page again.
Set -c 8192. The default context reserves about 5.5 GB of RAM; 8192 is enough for a full page and brings the total to about 1.6 GB.

The model may begin its answer with an empty <think></think> block: drop it.

On your phone, as an app

nuqta-ocr/apps has ready-made Android and iPhone apps that download these files once and read pages fully offline (first words in about 2 seconds with the 0.8B). They handle the input size, read the page in strips and use the right settings for you.

In chat apps (PocketPal, LM Studio, Jan, Ollama…)

These files carry Nuqta's reading settings, so llama.cpp-based apps pick them up without any setup: greedy decoding (temperature 0, no repeat penalty), thinking switched off in the chat template, and a minimum image size of about 1 megapixel so small pages and photos are enlarged before reading. With no settings at all, this raised a small page from 92.9% to 98.4% of characters right in our test, and stopped the model from writing <think>.

Some apps override these with their own defaults. If the output looks wrong, set them by hand:

Setting	Value
Temperature	0
Top-K	1
Repeat / frequency / presence penalty	off (1.0 / 0 / 0)
Context length	8192
Max new tokens	4096
System prompt	none
Thinking / reasoning	off

Send the image together with one of the prompts below in the same message. For the best accuracy on whole pages, use the nuqta-ocr library, which also applies the recommended input size.

Prompts

Mode	Prompt
as written (default)	`Transcribe the Arabic text in this image exactly as written. Include only the diacritics (tashkeel) that are visible in the image.`
no tashkeel	`Transcribe the Arabic text in this image without any diacritics.`
Markdown (tables, headings)	`Convert this document to Markdown: headings, paragraphs, lists and tables in reading order. Include only the diacritics visible in the image.`

Speed

Measured on a laptop CPU (8 threads, no GPU) with Q4_K_M, for one full book page:

Page size	Image encoding	Total
2.4 MP	18 s	39 s
1.2 MP	6 s	21 s

A text line takes about 2 s. A phone CPU is slower than this. A phone GPU (Metal on iPhone, Vulkan or OpenCL on Android) is usually much faster, especially for the image encoding step.

Limitations

Tables: the 0.8B may reorder or drop columns, especially in small mixed Arabic / English tables that start with an English column (every cell is still read correctly). Use the 4B GGUF for tables. Full limitations: Nuqta-Arabic-OCR-0.8B.

Contributors

Salah AlHaismawi · Mohamed Saleh Tulib · Ahmed ElSayed Mohamed Ibrahim

License

Apache 2.0, like the original model. Built on Qwen3.5-0.8B.

Citation

@misc{nuqta_ocr_2026,
  title  = {Nuqta OCR: Accurate Arabic Document OCR with Faithful Tashkeel},
  author = {Salah AlHaismawi and Mohamed Saleh Tulib and Ahmed ElSayed Mohamed Ibrahim},
  year   = {2026},
  url    = {https://huggingface.co/SalahALHaismawi/Nuqta-Arabic-OCR-0.8B-GGUF}
}

SalahALHaismawi/Nuqta-Arabic-OCR-0.8B-GGUF